Phikon-v2

At a glance

What it does

Tile embeddings evaluated mainly on predicting molecular biomarkers and clinical outcomes from H&E; successor to Phikon (ViT-B, TCGA-only).

Tasks, data types and cancers

Architecture

Training data

PANCAN-XL: 456M tiles from 58,359 WSIs across more than 30 cancer types; roughly half public (TCGA, CPTAC, GTEx, other public cohorts) and half private.

Linked datasets

• pretraining TCGA — The Cancer Genome Atlas (via NCI Genomic Data Commons) Free registration — part of PANCAN-XL — not a valid external test set
• pretraining CPTAC — Clinical Proteomic Tumor Analysis Consortium Free registration
• training TCGA — The Cancer Genome Atlas (via NCI Genomic Data Commons) Free registration
• training CPTAC — Clinical Proteomic Tumor Analysis Consortium Free registration

Evaluation

How to run

# generic timm loader — check the model card for the exact init args and image normalisation
import timm, torch
from huggingface_hub import login
login()  # gated repos: accept the licence on huggingface.co first
model = timm.create_model('hf-hub:owkin/phikon-v2', pretrained=True)
model.eval()
cfg = timm.data.resolve_data_config({}, model=model)
transform = timm.data.create_transform(**cfg)
# emb = model(transform(tile).unsqueeze(0))  # 1 x D tile embedding

Loads with AutoModel; note that TCGA and CPTAC are IN the pretraining set, so do not treat those as external test sets for this model.

Regulatory status and intended use

Source →

Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.

Limitations and bias

• Public benchmarks built on TCGA/CPTAC are contaminated for this model — evaluate on data outside PANCAN-XL.
• Non-commercial terms.

Sources

1. Filiot A et al. Phikon-v2, A large and public feature extractor for biomarker prediction. arXiv 2024
2. Hugging Face — owkin/phikon-v2